Re: [PATCH v3] hung_task: deduplicate identical hang reports

[David Laight <david.laight.linux@gmail.com>]

On Mon, 22 Jun 2026 17:58:54 +0200
Petr Mladek <pmladek@suse.com> wrote:

> On Sun 2026-06-21 17:37:56, Aaron Tomlin wrote:
> > Currently, during severe lock contention, multiple tasks can hang while
> > waiting on the exact same resource. The khungtaskd kthread
> > indiscriminately reports every single instance with a stack trace.
> > This can roll the kernel ring buffer and prematurely exhaust the
> > kernel.hung_task_warnings budget. Consequently, the kernel is left
> > entirely blind to subsequent, unrelated deadlocks.
> > 
> > To preserve the warning budget and ring buffer without sacrificing
> > observability, introduce a Wait Channel (wchan) and task-state based
> > deduplicator:
> > 
> >     1. Implement a lightweight, stack-allocated 64-slot Wait Channel
> >        (wchan) hash map. Tasks blocked on the exact same wchan during a
> >        single scan are recognised as sharing the same bottleneck,
> >        successfully deduplicating contentions even when the callers
> >        possess entirely disparate call stacks.  
> 
> I am sorry but I do not like this. It would show one random task blocked
> using a locking/wait API (mutex, semaphore, wait). But it will not be
> able to distinguish whether they are waiting for the same lock or
> event.
> 
> It might easily skip the lock/event which is the root of the problem.
> 
> By other words, the motivation for this patch is to avoid duplicated
> backtraces because the global limit of shown backtraces is too low
> and it hides too much. But this would hide even more backtraces.
> As a result administrators and developers will be even more blind.
> 
> Honestly, the previous version looked more acceptable to me. The
> problem with not-exactly same backtraces might be solved by
> comparing (hashing) only the top N backtrace levels, e.g. 10th.
> Anyway, we should compare the callers of the locking/waiter API.
> 
> IMHO, we should always print backtraces of all hung tasks when
> a hung_task is detected for the 1st time. Because we do not
> know which of the hung tasks is pointing to the root of the problem
> and which is a secondary victim.
> 
> Also I would primary try to increase the ring buffer size when
> backtraces get lost.

Mostly the traces wont be seen until they get written to file by syslogd
(assuming it can run).
So why not write them slowly enough that it keeps up?

> 
> >     2. Introduce a hung_task_reported bit-field in task_struct. If a task
> >        remains hung across multiple intervals, khungtaskd recognises it
> >        has already been reported. The bit is safely cleared without
> >        locks or atomics the moment the task's context switch counter
> >        increments.  
> 
> Also this looks like an interesting optimization which might help
> to reduce printing the same backtrace again and again. It looks
> much better than the global limit of printed backtraces.
> 
> >     3. For duplicate tasks, we still print the single-line
> >        "INFO: task ..." message and trigger tracepoint
> >        trace_sched_process_hang(). It merely skips calling
> >        sched_show_task() and debug_show_blocker(), printing a concise
> >        suppression notice instead.  
> 
> Yes, this is important as well.

And would need to include the pid of the duplicate trace so you can see
which one it is.

	David

> 
> Best Regards,
> Petr